Writing

What we build, test and stop, with dates, and one fresh paper at a time turned into a company note. Written for people who build AI products. RSS.

1. 2026-10-05We tested four lethal trifecta checkers. None caught the textbook caseWe ran four open-source lethal trifecta checkers on 180 tools from nine real MCP servers. All four missed local files plus fetch. The data and harness are public.
2. 2026-10-04The GitHub agent leak, and six patterns that block itIn May 2025 one public GitHub issue made an agent leak private repositories. Two weeks later its finders co-wrote six agent designs that block this attack.
3. 2026-10-03Why full power may waste energy in AI data centersA power aware scheduling service that recommends safe GPU caps for AI workloads, paid for by data center operators and managed AI providers.
4. 2026-09-04Evidence-native regulated operations: the company we would co-buildCorteus's working venture thesis. In regulated operations the costly step is proving an answer, not producing it. What a partner brings, and when we stop.
5. 2026-07-10What a real eval suite looks like, and how to ask for itA plain-English guide for investors and non-technical founders on what AI evals are, what a real suite contains, and how to tell evidence from theater.
6. 2026-07-06The Wrapper Test: will this AI product survive the next model release?An 8-question Wrapper Test for investors: how to tell whether an AI product has data, workflow depth, evals, and customers a model release cannot copy.